tensor-compiler/taco
Taco
Taco is a sparse tensor algebra compiler that translates high-level tensor index expressions into optimized machine code. It functions as a numerical code generator, producing specialized C kernels designed to execute complex multidimensional array operations efficiently on both CPU and GPU hardware.
The project distinguishes itself by allowing users to define custom tensor storage layouts by composing dimension-level formats, such as dense or compressed structures, to match the specific sparsity patterns of their datasets. By analyzing the mathematical structure of tensor operations at compile time, it automatically applies iteration-space transformations and kernel-level parallelization to minimize memory footprint and maximize execution speed.
Beyond its core compilation capabilities, the library supports a broad range of numerical processing tasks, including matrix-vector multiplication and various tensor contractions. It provides both a command-line interface for generating standalone kernels and a Python API for integrating high-performance tensor operations directly into numerical workflows.
Features
- High-Performance Tensor Libraries - Provides a high-performance compiler for executing complex multidimensional array operations with support for custom storage formats and parallel execution.
- Tensor Expression Compilers - Compiles complex tensor algebra expressions written in index notation into highly optimized machine code.
- Sparse Tensor Representations - Performs high-performance mathematical operations on multidimensional arrays that contain mostly zero values.
- Tensor Operation Compilers - Transforms complex mathematical index notation into highly optimized machine code for efficient execution on modern hardware.
- C Code Generators - Generates specialized C kernels from high-level tensor index expressions to accelerate numerical data processing on CPUs and GPUs.
- Tensor Contractions - Performs complex sparse and dense linear algebra calculations by compiling optimized kernels tailored to input data structure.
- Tensor Parallelism - Executes complex mathematical operations across multiple processor cores or graphics hardware to speed up heavy data processing.
- Iteration Space Optimizers - Optimizes data access patterns by automatically reordering loops and applying tiling strategies to match the underlying hardware memory hierarchy.
- Standalone Kernel Generators - Generates standalone C code for specific tensor algebra operations using a command-line interface.
- Parallel Processing - Executes intensive tensor algebra operations across multiple processor cores or graphics hardware to accelerate data analysis.
- Index Notation Parsers - Translates high-level tensor index expressions into optimized machine code by analyzing the mathematical structure of the operation.
- Table-Level Storage Formats - Allows defining custom tensor storage layouts by composing dimension-level formats to match specific data requirements.
- Specialized Storage Formats - Supports various storage formats such as CSR, COO, or CSF to reduce memory consumption and improve data access patterns.
- Sparsity-Aware Optimizers - Evaluates the sparsity patterns of input tensors at compile time to prune unnecessary computations and minimize memory footprint during execution.
- Tensor Computation Graphs - Applies advanced mathematical transformations like loop reordering and parallel splitting to ensure efficient data calculations.
- High-Performance Linear Algebra - Generates specialized high-performance C kernels for linear and tensor algebra tasks.
- Dimension-Level Formats - Constructs complex sparse tensor storage layouts by nesting primitive data structures like dense arrays and compressed fibers.
- Algebraic Lowering - Translates high-level tensor index expressions into optimized nested loop structures for efficient execution.
- Parallel Kernel Generators - Distributes tensor computations across multiple processor cores or graphics hardware by injecting synchronization primitives into the generated machine code.
Star history
How this analysis was created: This summary and feature list are AI-generated from collected project material and can contain mistakes. Stars, license and language are imported from GitHub. Inclusion does not mean that we have tested or audited this project. Check the source documentation for any feature you depend on. Learn more on our About page.
